A Novel Approach Combining Decoupling Training with Prototype-based Contrastive Learning for Tracking Long-Tailed Recording in Optical Aerial Imagery

Scene classification is a classical task in optical aerial imagery. In practical applications, it is difficult to obtain a balanced training samples supporting model training. Therefore, the dataset usually follow the Long-Tailed distribution, which is challenge for classification task, especially for multi-categories scene classification. It would make model intend to learn the representation of head classes with majority samples, which lead to under-represent middle and tail classes with minority samples and affect a better classifier construction for long-tailed classification. In this article, a novel Prototypical Supervised Contrastive Learning Based on Decoupling Framework called PSCL-DF is proposed for Long-Tailed Classification in Optical Remote Sensing Imagery to make a better feature learning and better classifier leaning. Finally, extensive tests are performed on the two artificially generated long-tailed datasets, demonstrating the robustness and effectiveness of the proposed methods. Our results show that decoupling feature extraction and classification achieve significant performance gains.